Supplementary Material for Experiments in "Quantitative comparison of deep learning-based image reconstruction methods for low-dose and sparse-angle CT applications"

Supplementing record containing (trained network) parameters of the reconstruction methods on the Apple CT Datasets in the article "Quantitative comparison of deep learning-based image reconstruction methods for low-dose and sparse-angle CT applications".

The experiments include 12 different settings:

• Noise settings: Noise-free, Gaussian noise, Scattering
• Numbers of angles: 50, 10, 5, 2

For each setting and each method, trained network parameters (or suitable hyper parameters for non-learned methods) are included.

Note: Parameters for the LoDoPaB-CT Dataset of those reconstructors implemented in DIVαℓ can be found in the supplementary repository supp.dival.

For details, see the article "Quantitative comparison of deep learning-based image reconstruction methods for low-dose and sparse-angle CT applications". See also the supplementary record containing saved test reconstructions and the supplementary repository providing source code. Below are references for the included methods.

• cinn: A. Denker et al., 2020, Conditional Normalizing Flows for Low-Dose Computed Tomography Image Reconstruction
• fbp: Filtered back-projection (ODL implementation)
• fbpistaunet: T. Liu et al., 2020, Interpreting U-Nets via Task-Driven Multiscale Dictionary Learning
• fbpmsdnet: D. Pelt et al., 2017, A mixed-scale dense convolutional neural network for image analysis
• fbpunet: K. H. Jin et al., 2017, Deep Convolutional Neural Network for Inverse Problems in Imaging
• ictnet: D. Bauer et al., 2021, iCTU-Net
• learnedpd: J. Adler et al., 2018, Learned Primal-Dual Reconstruction
• tv: Total Variation Regularization (DIVαℓ implementation)